Performance degradation with serialization of nested structs

[budrus]

The serialization performance at least for nested structs seems to be degraded with the latest cyclones-cxx.

Used commits:

• commits close to 0.9 release
• CycloneDDS b3c386e328fc2015bfa5494d6840be973fce2e7e
• CycloneDDS-CXX 725cc2487818268f351303e5bc3576f4fac89483

Tests that reproduce this:

• Open source performance test tool
  • Used branch

Expected behavior:

• Performance test setup: Struct32k, 1 pub, 1 sub, 100 Hz, BEST_EFFORT, VOLATILE, intraprocess, shmem disabled
• 100 samples are sent and received per second

Actual behavior:

• Only 1-2 samples are sent and received per second

[reicheratwork]

Looking in to this case.

[reicheratwork]

Can confirm this issue. It seems that the entity_properties_t meta data container is not that good at being recursively created, lots of creating/copying of std::list going on.

I am working on a different (a flat in stead of a tree-like) approach for storing the sample metadata.

[ErikAtApex]

I ran some more experiments. @reicheratwork Thanks for working on this; your fixes made a big difference. The performance for this deeply-nested structure is only 25% as fast as 0.8.1, but much better than the 1-2 samples per second.

• Environment:
  • Erik's laptop
  • Open source performance test tool
  • CycloneDDS-CXX docker container
  • Struct32k, 1 pub, 1 sub, 1000 Hz, BEST_EFFORT, VOLATILE
  • pub and sub are in the same process, but different threads
  • shmem disabled (CYCLONEDDS_URI environment variable is unset)
• Experiment 1:
  • CycloneDDS 0.8.1 tag
  • CycloneDDS-CXX 0.8.1 tag
  • 1ms mean latency
  • 1000 samples sent / received per second
• Experiment 2: (0.9ish, before this fix)
  • CycloneDDS b3c386e328fc2015bfa5494d6840be973fce2e7e
  • CycloneDDS-CXX 725cc2487818268f351303e5bc3576f4fac89483
  • at least 0.5s latency
  • 0 - 2 samples sent per second
  • We even tried an experiment with 0 subscribers, so it's just the publisher, and it still barely manages 2 samples per second
• Experiment 3: (0.9ish, after this fix)
  • CycloneDDS 13304f8e7256a24ed47e2b6505afa0dab7caaa75
  • CycloneDDS-CXX e954f150cbc7f67014b8a80e67c88c67f2afab01
  • 4ms mean latency
  • 250 samples sent / received per second